0
$\begingroup$

I am familiar with the log-rank test for comparing multiple Kaplan-Meier curves, but I am looking for a test that will show a trend between groups.

Lets say the groups are patients seen at different time periods (A 2000-2005; B 2005-2010; C 2010-2015).

As an example: we want to assess difference of mortality between time periods and maybe even trends according to the time periods. Lets say we have a mortality of 3%, 2% and 1% in time periods A, B and C, respectively. Should we just perform log rank pooled over strata or should we also check linear trend for factor levels in the log rank box in SPSS

Thank you very much in advance,

Kweet

$\endgroup$
  • $\begingroup$ Could you please say how you define time=0 for the individual cases when you prepare the Kaplan-Meier curves? Are you data amenable to Cox proportional hazards regression? $\endgroup$ – EdM Sep 30 at 14:13
  • $\begingroup$ Sure time = 0 means the start of the follow-up and time =365.25 is 1 year after start follow-up. Having an event is defined as 2, not having an event is defined as 1. Event is defined as mortality. It is amenable to Cox proportional hazards regression in terms of the assumption of proportionality is not violated. $\endgroup$ – Kweetvannix Sep 30 at 14:47
0
$\begingroup$

The first thing I would recommend would be to see whether any covariates potentially related to outcome changed over the time course from 2000 to 2015. For example, was some new type of therapy introduced over that time, or did measures of performance status of newly enrolled patients change over time? That will given an idea whether the differences in mortality are with respect to time of study entry per se or represent a change over time in characteristics of the patients enrolling in the study.

Then you can include the year (or time period) of entry into the study as a covariate in a Cox proportional hazards regression, along with standard clinical covariates. That will help control for covariates that might have changed in the enrolling population from 2000 to 2015. It might make sense to model year of entry as a continuous predictor, modeled flexibly with restricted cubic splines, rather than breaking year of entry up into arbitrary groups. See this page among others on this site for why binning a continuous predictor can be problematic.

Provided that the proportional hazards assumption is adequately respected, a significant coefficient for year (or time period) of entry would indicate a significant relationship between year of study entry and mortality. Furthermore, such a result with a Cox model would document that relationship while controlling for the other covariates.

$\endgroup$
  • $\begingroup$ Thank you very much for this comment! Some covariates did as you specified change over time. But the point is to not make a Cox regression/model, and just show crude data over time, albeit population does change over time, but that is another discussion point. Lets say the mortality decreased from 3% to 2% to 1% in period A, B and C, respectively. While the amount of comorbidities increased. We are focussing on the trend of mortality and discussing why there could be a trend. We just want to show crude data. $\endgroup$ – Kweetvannix Sep 30 at 20:05
  • $\begingroup$ @Kweetvannix if the proportional hazards assumption holds then the Cox model provides a simple solution: you plot the crude survival data for each of your 3 entry-time periods and use entry-time period, a 3-level categorical variable, as the sole predictor in the Cox model to quantify the differences. The overall significance test of the Cox model would provide a test of whether the 3 entry-time groups differ at all in terms of mortality, and you get hazard ratios too. Your study will be of more interest, however, if you can document from your data why mortality might have changed. $\endgroup$ – EdM Sep 30 at 20:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.